← Back to the main explorer · neighborhood context · outcomes over time · admitted vs. denied

The long view — admission chances and school achievement, 1999–2025

For each UC campus and each graduating class since 1999, one question: across California public high schools, did a school’s achievement level predict that class’s chance of admission? Two independent measurement families answer it — the college-bound tests (SAT, ACT), taken by self-selected testers through 2020, and the state’s census tests (STAR, CAHSEE, CAASPP), taken by every enrolled student, through 2025. Every point is dated by the graduating class it describes, so each test is matched to the students who actually sat it — a grade-11 test in spring 2019 describes the class of 2020. Where the two families overlap they agree, which is what permits reading the series as one arc. The association was positive for two decades, began to decay around 2016 — before the test-blind decision — and turns negative only after 2020, and only at some campuses.

Campus

Admit rate × school achievement, year by year

Each point is one graduating class: the correlation (Pearson r), across California public high schools with at least 25 applicants to the campus, between each school’s admit rate for that class and its achievement level, each instrument taken in the spring that class actually sat it. Points above zero mean higher-achieving schools saw higher admit rates that year; points below zero mean they saw lower ones. Solid census points sit on the era’s main grade-11 instrument (STAR CST, then CAASPP); hollow points sit on an instrument with known top-end compression (Stanford 9, CAHSEE) and read attenuated. Small open marks are a second census instrument describing the same class — where both exist they corroborate each other closely. Dashed connections mark instrument splices, not measured change. Hover any point for its instrument, its test spring and the exact value.

Table — every value plotted for this campus

The same series at every campus

Ten campuses, one axis

Identical scales throughout, graduating class on every axis; click a panel to load it above. The post-2020 inversion is confined to the more selective campuses.

Application volume by school advantage

UC applicants per 100 seniors, 1999–2025

A fixed panel of 757 schools observed in both 1999–2003 and 2021–2025, assigned once to five advantage quintiles on their long-run share of low-income, English-learner and foster students (UPP) — Q1 lowest need, Q5 highest. Lines are ratio-of-sums within quintile: total UC applicants per 100 seniors. The denominator is CDE grade-12 enrollment through 2016 and the ACGR cohort after (splice validated at r ≈ .98). The gap by advantage is at least a quarter-century old and has been narrowing in relative terms; the narrowing stalls after about 2013.

Advanced Placement expansion

AP exams per 100 seniors, 1999–2020

The same fixed-quintile design applied to the AP series — its own fixed panel of schools with AP data at both ends of the window (by coincidence also 757; it shares 745 schools with the application panel above but is not the same set). Counts are exams, not students — one student may take several — and exam totals after 2005–06 are derived from the published score distribution. The second view shows what share of those exams scored 3 or higher: participation multiplied severalfold everywhere, while the score mix rose slightly at the most advantaged schools and fell at the highest-need ones.

The gap between a school’s testers and its UC entrants

Enrollee SAT reading minus the school’s all-tester average, 2006–2015

For every school-year where both are published: the mean SAT critical-reading score of the students who enrolled at UC from a school, minus the mean of all SAT takers at that school the same year. The line is the yearly mean, the band the interquartile range across schools. UC’s intake sat well above its schools’ testing populations everywhere — positive in of school-years, overall mean + points — and the margin widened over the decade. It is wider at lower-achieving schools (r = with the school’s later CAASPP level).

Applicant grades against a flat yardstick

Applicant GPA drift, 2010–2016 — and what it is not

Each dot is a school in the Berkeley applicant pool with at least five observed years of both series. The left panel plots each school’s applicant-GPA trend against its SAT trend over the same years; in the shaded band the external yardstick was flat, yet applicant GPA still rose at of those schools (mean GPA points/year). The climb predates the CAASPP era and 2020. The right panel checks the natural suspect: schools’ AP expansion is essentially unrelated to their grade climb (r = ), and schools whose grades rose carry the same climb whether their AP programs expanded or stayed flat (index ) — so course-mix expansion does not account for it.

GPA trend vs. SAT trend, per school

Shaded band: |SAT trend| ≤ 0.02 z/year (“flat yardstick”). Hover a dot for the school.

Grade climb vs. AP expansion, per school

Inflation index (2015–19 vs. 2008–12) against the change in AP exams per senior. Hover a dot.

Methods — two rulers, one constant; and every caveat that governs this page

The page rests on two families of achievement measurement that are different constructs, not one construct with noise. The college-bound family (SAT, ACT, AP) is measured on self-selected testers — the college-bound core of a school. The census family (STAR, CAHSEE, CAASPP) is measured on essentially all enrolled students. Their school-level correlation is remarkably stable:

Across fifteen pre-2015 years the census and college-bound measures correlate at a mean of r = across schools (STAR-CST years alone: ), and in 2015–19 SAT × CAASPP runs .78–.81 — the same constant, two decades apart, on different instruments. The dips are instrument artifacts: the CAHSEE years (2002, 2014–15) sit lower because CAHSEE is a graduation bar that compresses the top of the scale, and 2002 additionally reflects its early administration.

Dating: every point is the graduating class, not the testing year. A test measures the students who sat it, and those students reach UC in a known year: the class of Y sits grade 9 in spring Y−3, grade 10 in spring Y−2, grade 11 in spring Y−1 and grade 12 in spring Y, then enters in fall Y. Each instrument is therefore shifted from its own test spring to the class it describes, so a point at 2020 always means the class of 2020 whichever instrument drew it.

Two consequences worth stating. First, this is the same convention the rest of this site uses, so the census gap sits in the same place on every page. Second, it is checkable: in the years when two different census instruments — different grades, different springs, different scales — both describe the same graduating class, they land in nearly the same place, a mean absolute difference of in r across campus-class pairs. Those second instruments are the small open marks on the chart. The alignment was not chosen to produce that agreement; the agreement is what the alignment buys.

Attenuation, and which way it points. With a between-family correlation of ρ ≈ .79, a correlation measured on the census ruler reads about 21% closer to zero than the same association measured on the college-bound ruler. All post-2020 points are census-ruler points (CAASPP is the only measure left after test-blind), so the post-2020 segment of the arc is understated, not inflated, relative to the pre-2020 college-bound segment — assuming the constant holds after 2020, which is untestable because the college-bound series ends there.
No pooling across scale eras. The SAT changed scale twice and its reporting basis twice more; STAR’s vocabulary changed in 2003; the census instrument changed twice. Every series here is built within (year × instrument) — correlations are computed within a single year on a single instrument, and raw levels are never compared across eras. Where a line crosses an instrument boundary the connection is drawn dashed and means only “the series continues,” never “this is measured change.”
CAHSEE’s ceiling, quantified. On the 2,487 schools present on both sides of the 2013 STAR→CAHSEE switch, the lowest decile moves +0.82 z and the top decile −0.30 z with no change in the schools — pure ceiling compression (strong schools pile up at 99–100% pass). CAHSEE-year correlations (hollow points) are therefore attenuated beyond the ordinary ρ ≈ .79. The 2001 CAHSEE sitting was a voluntary grade-9 administration and is excluded entirely.
Gaps are real gaps. On the census side the classes of 1999–2001 and 2003 have no instrument: grade-11 CST proficiency begins in spring 2003 (the class of 2004) and the grade-10 CAHSEE census begins in spring 2002 (the class of 2004), while the spring-2001 CAHSEE sitting was a voluntary grade-9 administration and is excluded throughout. The class of 2002 is carried only by Stanford 9. The classes of 2021 and 2022 have none: their grade-11 springs were 2020, cancelled, and 2021, excluded as non-representative. California also ended STAR after spring 2013 and ran spring 2014 only as a CAASPP field test with no reported results — that hole falls on the testing timeline, and on the class timeline it closes, because the class of 2014 is covered by grade-11 CST in spring 2013 and the class of 2015 by grade-10 CAHSEE in the same spring. The college-bound series ends with the class of 2020, the last one to apply with scores required. The 2019–20 SAT points sit on a grade-12-slice reporting basis (a population change, carried as its own era); San Diego’s 2020 college-bound point is anomalous under that change and should be read cautiously.
Santa Barbara diverges for real. At every other campus the two families agree up to the attenuation constant. Santa Barbara is the one campus with a genuine residual between the rulers (−0.117, every year 2015–19, with ACT siding against CAASPP as well) — there the two measures pull in opposite directions for reasons not established. Its panels carry this flag.
2020 is a one-year step, not part of the trend (ringed on the chart). The college-bound value moves sharply at most campuses in the 2020 cycle — up at San Diego, Santa Barbara and Irvine, down at Berkeley. It is not a measurement effect: the SAT reporting population changed a year earlier, at 2019, so it cannot produce a 2019–2020 step; the school measure itself is stable against a fixed external anchor (r = .77–.81 with a frozen spring-2019 census measure, every year 2016–2020); the step survives holding the achievement measure completely fixed; and the census family, which is an entirely separate instrument, shows the same step on the chart — at San Diego it runs −0.17 to +0.29 against the college-bound −0.19 to +0.30, and at Berkeley it turns down in the same year the college-bound line does. What moved was admission rates. The 2020 cycle was decided in March–April 2020, and campuses expanded admissions unevenly across schools: at San Diego the highest-achieving quartile of schools went from a 25.9% to a 35.0% mean admit rate while the lowest-achieving quartile fell from 30.9% to 28.6% — a reversal of that campus’s prior ordering within a single year. Berkeley moved the other way. Read 2020 as its own event; the surrounding trend is what the rest of the series shows.
Participation. The share of students taking the SAT (percent_tested) is published only through 2017–18; no participation-adjusted view is offered after that.
AP semantics. AP counts are exams, not students; totals after 2005–06 are derived from the published score distribution; the AP series is shown from 1999 on its own fixed panel.
What a point is, and is not. Every correlation is across schools (gate: ≥25 applicants to the campus; ≥40 such schools, else no point), between the school’s admit rate (admits ÷ applicants) and one achievement measure. It describes how a campus’s admission outcomes distribute over schools; it is not a statement about any individual applicant, and association is not attribution.
Counts universe. Admit rates are computed from the same published school × campus × year counts the rest of the explorer uses. The compilation behind the correlation series omits some very small schools (documented with the admitted-vs.-denied page); none of them passes the 25-applicant gate, and at the anchor point (Berkeley 2016) the full-universe counts give the identical result (r +0.50, n 440).
Advantage quintiles. UPP (unduplicated pupil percentage: the share of a school’s students who are low-income, English learner, or foster youth, counted once) averaged over all observed years, quintiled once — membership never changes, so trends cannot be group-switching artifacts.

Data vintage

Admissions: UC Information Center, freshman admissions by source high school (California public HS), admission years fall 1999–2025; school-year admit rates and applicant counts as compiled for the explorer’s school×year panel.
College-bound achievement: CDE SAT, ACT and AP research files by school, 1998–99 to 2019–20, compiled to one row per school×year with per-year rank-space bridges across the SAT’s scale eras; CDS-code defect in the 2012–13 (and ACT 2007–08) files repaired before use.
Census achievement: CDE STAR CST grade 11 ELA proficiency (springs 2003–2013) — the same construct CAASPP measures — with CDE CAHSEE grade 10 (springs 2002–2015), CAASPP grade 11 (springs 2015–2024; spring 2020 cancelled, spring 2021 excluded) and Stanford 9 (spring 2000, grades 9–11 pooled). Each value is dated by the graduating class that sat it, with its instrument and test spring recorded on every point.
Dating: Every series is dated by the normative graduating class, which is also the UC admission year: grade 11 tests shift +1 year from their spring, grade 10 tests +2, and the senior-weighted SAT/ACT files need no shift. The methods note gives the full table.
This page: Every series precomputed by build/build_history_data.py into history/data_history.js; no computation happens in the browser beyond drawing.

Last updated 1 August 2026. Revision history is in docs/CHANGELOG.md.